AI-based motion subject encounter probability determination and collision avoidance warning method and system

Abstract

This invention discloses an AI-based method and system for determining the probability of encounters between moving subjects and for collision avoidance warning. It relates to the interdisciplinary fields of artificial intelligence, probability statistics, trajectory prediction, and aerospace / intelligent transportation. This invention deeply integrates AI trajectory prediction with encounter probability calculation, effectively reducing the cost of manual intervention. By adding Gaussian noise to the trajectory distribution output by the AI ​​trajectory prediction model to quantify prediction uncertainty, the encounter probability calculation fully considers the statistical characteristics of prediction errors, improving calculation accuracy. The encounter probability calculation is based on a first position distribution and a second position distribution, avoiding the accuracy limitations of simple statistical methods. Collision avoidance control commands or warning signals are generated based on the encounter probability. The collision avoidance control commands are sent to the actuator to perform braking or avoidance operations, or the warning signals are sent to the display terminal or alarm device to alert the operator, thus enhancing the level of intelligence.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, probability and statistics, trajectory prediction and aerospace / intelligent transportation intersection technology, and in particular to an AI-based method and system for determining the probability of encounters between moving subjects and for collision avoidance warning. Background Technology

[0002] With the rapid development of aerospace, intelligent transportation, and other fields, the need for risk assessment of encounters involving multiple moving entities (satellites, vehicles, drones, etc.) is becoming increasingly urgent. In applications such as satellite collision warning, intelligent traffic management, and drone obstacle avoidance, it is necessary to calculate the probability of future encounters of multiple moving entities in real time to determine whether a collision risk exists and to execute warnings or proactive collision avoidance actions accordingly. However, in real-world scenarios, motion trajectories are affected by various factors such as environmental disturbances, sensor noise, and control command delays, resulting in significant uncertainty in trajectory prediction. How to accurately and efficiently calculate encounter probabilities under this uncertainty has become a core technical challenge that urgently needs to be addressed.

[0003] To address the aforementioned problem of calculating the intersection probability, several solutions have been proposed in existing technologies. One type of solution employs Monte Carlo simulation, which involves adding random perturbations to the trajectory of the moving subjects and sampling multiple times. The frequency with which the distance between samples is less than a safe distance threshold is used as the estimated intersection probability. This method is simple in principle, easy to implement, and applicable to various motion scenarios. Another type of solution uses analytical calculation. Assuming that the trajectory prediction error follows a Gaussian distribution, it establishes a joint Gaussian distribution model for the two moving subjects and integrates the joint probability density function within the spatial region limited by the safe distance to directly obtain the analytical expression for the intersection probability. These two methods represent two technical routes: statistical simulation and analytical calculation, respectively.

[0004] However, the existing technologies still have the following significant shortcomings. First, traditional calculation methods do not fully consider the uncertainty characteristics of the output of artificial intelligence trajectory prediction models. They only calculate the intersection probability based on a single predicted trajectory, which cannot effectively quantify the impact of prediction errors, resulting in large deviations in calculation results and making it difficult to adapt to complex real-world scenarios. Second, simple statistical methods have low accuracy, while high-precision analytical methods have strict requirements that trajectory errors must follow a Gaussian distribution. When the actual error distribution deviates from the Gaussian shape, the generalization ability is weak and cannot adapt to the diverse error characteristics under different scenarios. Third, it is difficult to balance computational efficiency and accuracy. Monte Carlo methods either have too few samples, resulting in insufficient accuracy, or too many samples, affecting real-time performance and failing to meet the needs of high-frequency scenarios. Fourth, when there are three or more moving entities, multi-entity intersection calculations involve a large number of pairwise combinations, which drastically increases computational complexity. Existing methods lack efficient joint probability calculation schemes and have poor multi-entity adaptability. Fifth, existing methods are not deeply integrated with artificial intelligence trajectory prediction models. They require manual processing of trajectory data as input, resulting in low automation and difficulty in embedding them into existing intelligent decision-making systems. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing technologies, specifically providing an AI-based method and system for determining the probability of collisions between moving subjects and for collision avoidance warning, as detailed below: 1) In a first aspect, the present invention provides an AI-based method for determining the probability of collision between moving subjects and for collision avoidance warning, the specific technical solution of which is as follows: Obtain the initial motion parameters of the first and second moving bodies; The initial motion parameters are input into a pre-trained AI trajectory prediction model to predict the trajectory distribution of the first and second moving subjects within the time window, respectively. Gaussian noise is added to the trajectory distribution to quantify the prediction uncertainty, thereby obtaining the first position distribution of the first moving subject and the second position distribution of the second moving subject. Based on the first position distribution and the second position distribution, the intersection probability of the first moving subject and the second moving subject is obtained; Based on the probability of encounter, a collision avoidance control command or a warning signal is generated, and the collision avoidance control command is sent to the actuator of the first moving body and / or the second moving body, or the warning signal is sent to the display terminal or alarm device.

[0006] The beneficial effects of the AI-based method for determining the probability of collision and providing early warning of collision avoidance for moving subjects provided by this invention are as follows: This system deeply integrates AI trajectory prediction with intersection probability calculation, directly using initial motion parameters as input. This avoids manual trajectory data processing, automating the entire process from trajectory prediction to collision avoidance command or warning signal generation. This effectively reduces the cost of manual intervention and facilitates integration into existing intelligent decision-making systems. By adding Gaussian noise to the trajectory distribution output by the AI ​​trajectory prediction model to quantify prediction uncertainty, the intersection probability calculation fully considers the statistical characteristics of prediction errors. This overcomes the bias caused by traditional methods based on single trajectory calculations, improving calculation accuracy and reducing the risk of misjudgment and missed judgment. The intersection probability calculation is based on the first and second position distributions, fully utilizing the distribution information output by AI trajectory prediction and avoiding the inaccuracies of simple statistical methods. Collision avoidance control commands or warning signals are generated based on the intersection probability. The collision avoidance control commands are sent to the actuators to perform braking or avoidance operations, or the warning signals are sent to the display terminal or alarm device to alert the operator. This provides clear intervention basis for the upper-level decision-making system and improves the overall intelligence level of the system.

[0007] Based on the above scheme, the AI-based method for determining the probability of collision of moving subjects and for collision avoidance warning can be further improved as follows.

[0008] Furthermore, based on the first and second position distributions, the intersection probability of the first and second moving subjects is obtained, including: when using the Monte Carlo simulation mode, multiple sets of position samples are collected from the first and second position distributions respectively, each time step is traversed, the distance between the samples of the first and second moving subjects at each time step is calculated, the number of intersections where the distance between the samples is less than or equal to the safe distance threshold is counted, and the number of intersections is divided by the product of the total number of samples and the total number of time steps to obtain the intersection probability; or, when using the Gaussian distribution analytical mode, the first Gaussian parameter of the first position distribution and the second Gaussian parameter of the second position distribution are extracted, the Gaussian distribution of the relative position is calculated based on the first and second Gaussian parameters, and the intersection probability is obtained by numerically integrating the Gaussian distribution of the relative position within the area limited by the safe distance threshold.

[0009] The advantages of adopting the above-mentioned further scheme are: it provides two switchable intersection probability calculation schemes: Monte Carlo simulation mode and Gaussian distribution analytical mode. The Monte Carlo simulation mode calculates the intersection probability by randomly sampling position samples and performing distance statistics, without requiring the error to follow a specific distribution, making it suitable for general scenarios. The Gaussian distribution analytical mode extracts Gaussian parameters and calculates the Gaussian distribution of relative positions, then performs numerical integration within a region limited by a safe distance threshold to obtain the intersection probability, making it suitable for high-precision scenarios. The two modes can be flexibly selected according to actual needs, enabling the method to adapt to application scenarios with different accuracy and real-time requirements, effectively improving the method's generalization ability and scenario adaptability.

[0010] Furthermore, when using the Monte Carlo simulation mode, the total number of samples is dynamically adjusted according to the preset accuracy requirements; and / or, when using the Gaussian distribution analytical mode, if the error of the first position distribution or the second position distribution does not conform to the Gaussian distribution, the kernel density is used to estimate the distribution of the fitting error.

[0011] The beneficial effects of adopting the above-mentioned further scheme are as follows: By dynamically adjusting the total number of samples, the number of samples is automatically increased or decreased according to preset accuracy requirements in Monte Carlo simulation mode. When high accuracy is required, the number of samples is increased to reduce statistical errors; when high real-time requirements are required, the number of samples is decreased to improve computational speed, achieving a flexible balance between computational accuracy and efficiency. In Gaussian distribution analytical mode, when the error does not follow a Gaussian distribution, kernel density estimation is used to fit the error distribution, replacing the original Gaussian noise generation method. This allows the positional distribution to adapt to error characteristics of any shape, avoiding computational deviations caused by mismatched distribution assumptions. These two optimization mechanisms together improve the computational reliability and adaptability of the method in different scenarios.

[0012] Furthermore, it also includes: for multi-moving-subject scenarios, decomposing the multi-moving-subject into multiple moving-subject pairs, performing initial motion parameter acquisition, trajectory distribution prediction, and intersection probability calculation for each moving-subject pair, and outputting the intersection probability of each moving-subject pair; or, merging and calculating the joint intersection probability of all moving-subjects through a preset hazard domain simplification strategy, generating collision avoidance control commands or warning signals based on the intersection probability or joint intersection probability of each moving-subject pair, sending the collision avoidance control commands to the actuators of the corresponding moving-subjects, or sending the warning signals to the display terminal or alarm device.

[0013] The advantages of adopting the above-mentioned further solutions are as follows: The moving subject pair decomposition method splits multiple moving subjects into multiple moving subject pairs. Each moving subject pair independently performs the intersection probability calculation and outputs the intersection probability of each moving subject pair, transforming the complex pairwise combination problem into independent pairing calculations. The hazard domain simplification strategy avoids the tedious process of calculating the joint intersection probability of all moving subjects by merging the calculations of the joint intersection probability of all moving subjects, further reducing computational complexity. Both strategies can generate collision avoidance control commands or warning signals based on the calculation results, and send the collision avoidance control commands to the actuators of the corresponding moving subjects, or send the warning signals to the display terminal or alarm device, providing an efficient and practical calculation solution for intersection risk assessment in multi-moving subject scenarios.

[0014] Furthermore, the joint intersection probability of all moving subjects is calculated by combining the pre-set danger domain simplification strategy, including: determining the mean point of the position distribution of all moving subjects at each time step within the time window, calculating the convex hull boundary of all mean points, expanding the convex hull boundary outward with a safety distance threshold to obtain the danger domain, and then calculating the ratio of the number of position samples of each moving subject falling into the danger domain to the total number of samples, and taking the maximum value of the ratio of all moving subjects as the joint intersection probability.

[0015] The beneficial effects of adopting the above-mentioned further scheme are as follows: First, determine the mean point of the position distribution of all moving subjects at each time step, calculate the convex hull boundary of all mean points, and then extend the convex hull boundary outward by a safety distance threshold to obtain the danger domain. This ensures that the danger domain spatially covers all moving subjects and their safety interval range, eliminating the need for pairing and calculating moving subjects one by one, significantly reducing the computational complexity of multi-moving-subject intersection probability. Then, calculate the ratio of the number of position samples of each moving subject falling into the danger domain to the total number of samples. The maximum ratio of all moving subjects is taken as the joint intersection probability, reflecting the highest intersection risk level faced by any moving subject in the entire scene with a single value. This provides a clear basis for subsequent overall risk assessment and reduces the computational resource consumption in multi-moving-subject scenarios.

[0016] Furthermore, generating a collision avoidance control command or a warning signal based on the meeting probability includes: comparing the meeting probability with a preset first threshold and a second threshold, wherein the first threshold is greater than the second threshold; when the meeting probability is greater than or equal to the first threshold, generating an emergency braking command as a collision avoidance control command, wherein the emergency braking command instructs the actuators of the first moving body and / or the second moving body to perform an emergency braking operation; when the meeting probability is less than the first threshold but greater than or equal to the second threshold, generating a deceleration and avoidance command as a collision avoidance control command, or generating a warning signal containing risk location and risk time information and sending it to a display terminal or alarm device.

[0017] The beneficial effects of adopting the above-mentioned further scheme are as follows: By using a preset first threshold and a second threshold, the probability of collision is divided into different risk levels, with the first threshold being greater than the second threshold. When the probability of collision is higher than or equal to the first threshold, it is determined to be high risk, generating an emergency braking command to instruct the actuator to perform an emergency braking operation with maximum braking force, achieving mandatory intervention. When the probability of collision is between the first and second thresholds, it is determined to be medium risk, generating a deceleration and avoidance command or a warning signal containing information on the risk location and risk time. This can guide the moving entity to perform appropriate avoidance and also prompt the operator for manual intervention. By adopting differentiated response strategies for different levels of collision risk through a hierarchical decision-making mechanism, false triggering or missed triggering caused by a single threshold judgment is avoided, improving the rationality and effectiveness of collision avoidance decisions.

[0018] 2) In a second aspect, the present invention also provides an AI-based system for determining the probability of collisions between moving subjects and for collision avoidance warning, the specific technical solution of which is as follows: It includes a motion parameter acquisition module, a prediction addition module, a intersection probability calculation module, and a collision avoidance warning module; The motion parameter acquisition module is used to: acquire the initial motion parameters of the first moving body and the second moving body; The prediction addition module is used to: input the initial motion parameters into the pre-trained AI trajectory prediction model, predict the trajectory distribution of the first moving subject and the second moving subject within the time window respectively, and add Gaussian noise to the trajectory distribution to quantify the prediction uncertainty, so as to obtain the first position distribution of the first moving subject and the second position distribution of the second moving subject. The intersection probability calculation module is used to: obtain the intersection probability of the first moving subject and the second moving subject based on the first position distribution and the second position distribution; The collision avoidance warning module is used to: generate collision avoidance control commands or generate warning signals based on the probability of encounter, and send the collision avoidance control commands to the actuators of the first moving body and / or the second moving body, or send the warning signals to the display terminal or alarm device.

[0019] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor, so that the electronic device can realize any of the above-mentioned AI-based methods for determining the probability of intersection of moving subjects and collision avoidance warning.

[0020] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements any of the above-mentioned AI-based methods for determining the probability of intersection of moving subjects and for collision avoidance warning.

[0021] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below: Figure 1 This is a flowchart illustrating an AI-based method for determining the probability of collisions between moving subjects and for collision avoidance warning, according to an embodiment of the present invention. Figure 2This is a schematic diagram of an AI-based system for determining the probability of collisions between moving subjects and for collision avoidance warning, according to an embodiment of the present invention. Detailed Implementation

[0023] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0024] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0025] like Figure 1 As shown in the figure, an AI-based method for determining the probability of collision and providing collision avoidance warning for moving subjects according to an embodiment of the present invention includes the following steps: S1. Obtain the initial motion parameters of the first and second moving subjects. The specific implementation process is as follows: S10. Determine the contents of the initial motion parameters. The initial motion parameters include at least the position, velocity, and direction of motion of the first and second moving bodies at the initial moment. Position is used to determine the coordinates of the moving body in space, velocity is used to describe the magnitude of the displacement of the moving body per unit time, and direction of motion is used to describe the direction of the displacement of the moving body.

[0026] S11. For spacecraft such as satellites, their position, velocity, and direction of motion can be obtained through a Global Navigation Satellite System (GNSS) receiver. For moving objects such as vehicles and drones, their position, velocity, and direction of motion can be obtained through a GNSS receiver, inertial measurement unit (IMU), radar, or optical sensors. The GNSS receiver calculates the three-dimensional position and velocity of the moving object by receiving navigation satellite signals. The IMU measures the acceleration and angular velocity of the moving object and obtains velocity and position information through integration. The radar transmits electromagnetic waves and receives target echoes to measure the distance, azimuth, and radial velocity of the moving object. The optical sensor uses image acquisition and target recognition algorithms to calculate the position offset of the moving object in the image and obtains its actual spatial position by combining camera parameters. When a single sensor cannot meet the accuracy requirements, a multi-sensor data fusion method can be used to fuse measurement data from two or more sensors, including the GNSS, IMU, radar, and optical sensors. After processing by filtering algorithms such as Kalman filtering or unscented Kalman filtering, optimized estimates of position, velocity, and direction of motion are output.

[0027] S12. Convert the raw data collected by each sensor to the same coordinate system to eliminate deviations caused by different sensor installation positions or coordinate system definitions. Timestamp align the position coordinates output by the GNSS receiver, the acceleration and angular velocity output by the inertial measurement unit, the range and azimuth output by the radar, and the image coordinates output by the optical sensor to ensure all parameters correspond to the same starting time. After preprocessing, the initial position, initial velocity, and initial direction of motion of the first and second moving subjects in the unified coordinate system are obtained, serving as input for the subsequent AI trajectory prediction model.

[0028] S2. Input the initial motion parameters into the pre-trained AI trajectory prediction model to predict the trajectory distributions of the first and second moving subjects within the time window, respectively. Add Gaussian noise to the trajectory distributions to quantify the prediction uncertainty, and obtain the first position distribution of the first moving subject and the second position distribution of the second moving subject. The specific implementation process is as follows: S20. Constructing the network structure of the AI ​​trajectory prediction model. The AI ​​trajectory prediction model adopts a Long Short-Term Memory (LSTM) network, a variant of recurrent neural networks that excels at processing time-series data. Through three gating structures—input gate, forget gate, and output gate—it controls the flow and memory of information, effectively capturing the long-term dependency features of a moving subject's trajectory in the time dimension. The network structure of the AI ​​trajectory prediction model consists of an input layer, an LSM layer, and a fully connected layer stacked sequentially. The input layer receives an initial sequence of motion parameters arranged in time steps. The input data for each time step includes values ​​for the moving subject in three dimensions: position, velocity, and direction of motion. The LSM layer is configured as two layers: the first LSM layer contains 128 hidden units, and the second LSM layer contains 64 hidden units. Both layers employ a bidirectional LSM structure, which processes the input sequence simultaneously in both forward and backward directions, enabling the model to utilize contextual information from past and future moments for trajectory prediction. Each LSM layer is followed by a dropout layer. During training, the dropout layer randomly sets the output of some hidden units to zero with a preset dropout probability to prevent overfitting. The fully connected layer maps the hidden state output by the long short-term memory layer to the position coordinates of the predicted trajectory. The output dimension of the fully connected layer is 2, which corresponds to the horizontal and vertical coordinates of the predicted position in the planar coordinate system.

[0029] S21. Collect a large amount of real motion trajectory data of moving subjects within historical time periods as a training dataset. Each trajectory data includes the position, velocity, and direction of motion of the moving subject in a continuous time series. Clean the collected raw trajectory data, removing abnormal and missing data points caused by sensor failure or communication interruption. Perform a sliding window segmentation operation on the cleaned trajectory data, setting the window length to the number of historical time steps plus the number of predicted time steps. Slide the window along the time axis with a fixed step size to extract multiple trajectory segments from each complete trajectory. The previous historical time step data of each trajectory segment is used as the model input feature, and the subsequent predicted time step data is used as the model output label. Normalize the position coordinates, velocity, and direction of motion in all trajectory segments, scaling the values ​​of each dimension to a standard range with a mean of zero and a variance of one, eliminating the influence of different units on model training. Divide the normalized trajectory segments into training, validation, and test sets according to a preset ratio. The training set is used to update model parameters, the validation set is used to monitor the model's generalization ability and adjust hyperparameters during training, and the test set is used to finally evaluate the model's prediction performance.

[0030] S22. A supervised learning approach is used to train the AI ​​trajectory prediction model. The initial motion parameter sequence of historical time steps in the trajectory segment is used as the model input, and the actual trajectory position coordinates of the corresponding prediction time step are used as supervision labels. Mean squared error is selected as the loss function. Mean squared error measures the average of the squared differences between the model's predicted position and the actual position. Its calculation formula is: ,in, This represents the value of the loss function. This indicates the number of trajectory segment samples contained in a training batch. This represents the total number of prediction time steps. Indicates the first The sample at the th The predicted x-axis value for each prediction time step. Indicates the first The sample at the th The true value of the x-axis at each predicted time step. Indicates the first The sample at the th The predicted value of the ordinate at each prediction time step. Indicates the first The sample at the th The true values ​​of the ordinate at each prediction time step are used. An adaptive moment estimation optimizer (IME) is selected as the parameter update algorithm. The IME adaptively adjusts the learning rate of each parameter by calculating the first and second moment estimates of the gradient. During training, training set data is input into the AI ​​trajectory prediction model in batches. Forward propagation calculates the loss function value, and backpropagation over time calculates the gradient of the loss function with respect to each trainable parameter in the model. The IME updates the model parameters based on the gradient. After each training epoch, validation set data is input into the model to calculate the loss function value, and the model's performance on the validation set is monitored. When the validation set loss function value no longer decreases in multiple consecutive training epochs, training is terminated early to prevent overfitting. After training, the model parameters with the smallest loss function value on the validation set are selected as the final AI trajectory prediction model parameters.

[0031] S23. Arrange the initial position, initial velocity, and initial direction of motion of the first moving subject in chronological order to form a historical time step sequence. Input this sequence into the trained AI trajectory prediction model. After forward propagation, the model outputs the average predicted position of the first moving subject at each predicted time step within the time window. Similarly, input the initial motion parameters of the second moving subject into the same AI trajectory prediction model, and output the average predicted position of the second moving subject at each predicted time step within the time window. The length of the time window is equal to the total number of predicted time steps. Multiply by the time interval of each time step; the value of the time window is preset according to the specific application scenario.

[0032] S24. The mean value of the predicted position output by the AI ​​trajectory prediction model at each prediction time step is used as the mean parameter of the Gaussian distribution. A preset noise standard deviation is set as the standard deviation parameter of the Gaussian distribution. The specific value of the standard deviation is determined based on the prediction error statistics of the AI ​​trajectory prediction model on the test set. Gaussian noise with a mean of zero and a standard deviation equal to the preset noise standard deviation is added to the mean value of the predicted position of the first moving subject at each prediction time step to generate a position sample set of the first moving subject at that time step. The same Gaussian noise is added to the mean value of the predicted position of the second moving subject at each prediction time step to generate a position sample set of the second moving subject at that time step. After the above processing, the first position distribution of the first moving subject at each prediction time step within the time window and the second position distribution of the second moving subject at each prediction time step within the time window are obtained. The first position distribution and the second position distribution are represented in the form of position sample sets, and the number of samples in each position sample set is determined by the Monte Carlo simulation mode or Gaussian distribution analytical mode used in the subsequent intersection probability calculation.

[0033] S3. Based on the first position distribution and the second position distribution, the intersection probability of the first moving subject and the second moving subject is obtained, which is achieved in the following way: (1) First implementation method: When using Monte Carlo simulation mode, multiple sets of position samples are collected from the first position distribution and the second position distribution respectively. Each time step is traversed, and the distance between the samples of the first moving subject and the second moving subject at each time step is calculated. The number of intersections where the distance between the samples is less than or equal to the safe distance threshold is counted. The number of intersections is divided by the product of the total number of samples and the total number of time steps to obtain the intersection probability. The specific implementation process is as follows: 1) Determine the total number of samples and perform position sample acquisition. The total number of samples is preset according to the accuracy requirements; the higher the accuracy requirements, the larger the total number of samples. From the position sample set distributed in each time step at the first position, a number of position samples equal to the total number of samples are collected by random sampling. Simultaneously, the same number of position samples are collected from the position sample set distributed in the corresponding time step at the second position, ensuring that the first and second moving subjects each have the total number of position samples for subsequent calculations at each time step.

[0034] 2) Iterate through all time steps within the time window and calculate the inter-sample distance for each time step. For each time step currently being iterated, match the total number of position samples of the first moving subject with the total number of position samples of the second moving subject according to the sample index, forming a total number of sample pairs. For each sample pair, calculate the inter-sample distance between the position samples of the first moving subject and the position samples of the second moving subject. The inter-sample distance uses Euclidean distance, which represents the straight-line distance between two points in a planar coordinate system. The formula for calculating the inter-sample distance is: ,in, Indicates the index of the time step. The value range is from the first time step to the total number of time steps. , The value is equal to the total number of discrete time points within the time window; Indicates the index of the sample pair. The value range is from the first sample pair to the total number of samples. , Indicates the total number of samples; Indicates the first moving subject in the th The first time step The x-axis of each sample; Indicates the first moving subject in the th The first time step The ordinate of each sample; Indicates the second moving subject in the first... The first time step The x-axis of each sample; Indicates the second moving subject in the first... The first time step The ordinate of each sample; Indicates the first The first time step The inter-sample distance of each sample pair is calculated using the formula described above.

[0035] 3) Count the number of intersections where the distance between samples is less than or equal to the safe distance threshold. The safe distance threshold is a pre-set distance limit used to determine whether two moving objects have intersected. When the actual distance between two moving objects is less than or equal to the safe distance threshold, the two moving objects are considered to have intersected at that moment. The calculated distance between each sample... With safe distance threshold Comparison, Indicates the safe distance threshold. When When a crossover event occurs, the sample pair is considered a crossover event, and the number of crossovers is incremented. After traversing all time steps and all sample pairs, the total number of crossovers is obtained. , This indicates the total number of intersections.

[0036] 4) The total number of intersections obtained Divide by the total number of samples Total number of time steps The product of these factors yields the intersection probability. The formula for calculating the intersection probability is: ,in, Represents the probability of intersection. This represents the total number of intersections. Represents the total number of samples. This represents the total number of time steps. Total number of samples. Total number of time steps The product represents the total number of all sample pairs at all time steps, and the intersection probability. The value is equal to the proportion of sample pairs that intersect among all sample pairs. After the above steps, the intersection probability of the first moving subject and the second moving subject within the time window is obtained.

[0037] When using the Monte Carlo simulation mode, the total number of samples is dynamically adjusted according to the preset accuracy requirements. Specifically: 1) Determine the mapping relationship between accuracy requirements and the total number of samples. The preset accuracy requirement is quantified by the relative error limit or confidence interval width of the intersection probability estimate. The smaller the relative error limit or the narrower the confidence interval, the higher the accuracy requirement. According to the statistical principles of Monte Carlo simulation, the standard deviation of the intersection probability estimate is inversely proportional to the square root of the total number of samples. The larger the total number of samples, the smaller the statistical fluctuation of the estimate and the higher the calculation accuracy. Pre-set a minimum and a maximum value for the total number of samples. The minimum value ensures basic calculation accuracy, while the maximum value limits the upper limit of computational resource consumption. The minimum value is set to 5,000, and the maximum value is set to 20,000.

[0038] 2) When starting the Monte Carlo simulation mode, the total number of samples is set to a preset default value, which is the midpoint between the minimum and maximum values. This default value serves as the starting point for dynamic adjustments, and subsequent adjustments are made based on the actual satisfaction of accuracy requirements.

[0039] 3) Calculate the statistical error index of the intersection probability estimates under the current total number of samples. In a complete Monte Carlo simulation, all sample pairs are randomly divided into multiple equal-sized sub-batches. Each sub-batch independently calculates its intersection probability estimate, resulting in multiple intersection probability estimates. The standard deviation of these intersection probability estimates is calculated as the statistical error index. A smaller standard deviation indicates a more stable calculation result and higher accuracy under the current total number of samples.

[0040] 4) When the statistical error index is less than or equal to the preset allowable error value corresponding to the accuracy requirement, keep the current total number of samples unchanged, and consider the current accuracy to meet the requirement. When the statistical error index is greater than the preset allowable error value corresponding to the accuracy requirement, increase the total number of samples by one step, with the step size being one-tenth of the difference between the minimum and maximum values. The increased total number of samples should not exceed the preset maximum value. After increasing the total number of samples, re-execute the Monte Carlo simulation calculation and recalculate the statistical error index. Repeat the above comparison and adjustment process until the statistical error index meets the preset accuracy requirement or the total number of samples reaches the preset maximum value.

[0041] (2) Second implementation method: When using the Gaussian distribution analytical mode, extract the first Gaussian parameter of the first position distribution and the second Gaussian parameter of the second position distribution. Calculate the Gaussian distribution of the relative position based on the first and second Gaussian parameters. Obtain the intersection probability by numerically integrating the Gaussian distribution of the relative position within the area limited by the safety distance threshold. The specific implementation process is as follows: 1) Extract the first Gaussian parameters of the first positional distribution and the second Gaussian parameters of the second positional distribution. The first and second positional distributions are each represented as Gaussian distributions, each containing a mean vector and a covariance matrix. The first Gaussian parameters include the first mean vector and the first covariance matrix, where the first mean vector is denoted as... , ,in This represents the mean of the predicted x-coordinate of the first moving subject at a certain time step within the time window. The mean of the ordinate is represented by ; the first covariance matrix is ​​denoted as . , It is a two-dimensional square matrix, whose diagonal elements are the variances of the x-coordinates and y-coordinates, respectively, and whose off-diagonal elements are the covariances of the x-coordinates and y-coordinates. The second Gaussian parameters include the second mean vector. Second covariance matrix , , This represents the mean of the x-coordinate of the predicted position of the second moving subject at the corresponding time step. This represents the mean of the ordinate. Structure and Same. Read directly from the first position distribution. and The first Gaussian parameter is directly read from the second position distribution. and As the second Gaussian parameter.

[0042] 2) Calculate the Gaussian distribution of the relative positions based on the first and second Gaussian parameters. The relative position is defined as the difference between the positions of the first and second moving subjects. The mean vector of the relative positions is also calculated. It equals the first mean vector minus the second mean vector, calculated using the following formula: Covariance matrix of relative positions It equals the sum of the first and second covariance matrices, calculated as follows: After the above calculations, the Gaussian distribution of the relative positions is obtained, and the mean vector of this Gaussian distribution is... The covariance matrix is The Gaussian distribution of relative positions is used to describe the uncertainty of the relative positions between two moving entities.

[0043] 3) Numerically integrate the Gaussian distribution of the relative positions, and obtain the intersection probability by integrating within the region defined by the safety distance threshold. The safety distance threshold is denoted as... , This represents the maximum distance limit at which two moving objects will intersect. Intersection probability. The probability density function equal to the relative position satisfies the condition that the distance is less than or equal to The integral over the entire relative position region is expressed as: ,in, and These represent the two coordinate components of the relative position in the planar coordinate system. The two-dimensional Gaussian probability density function representing relative position is given by... and Deterministic numerical integration is employed, using deterministic numerical integration methods, where the integration region satisfies... The circular region is divided into equally spaced rectangular grids on a two-dimensional plane, and the side length of each grid is denoted as . , The value is set according to the preset integration accuracy requirements. The entire grid is traversed, and for each grid cell, the probability density value at the grid center point is calculated. Multiply this value by the grid area The contribution of each grid cell to the integration is obtained. The contributions of all grid cells are summed, and the sum is the intersection probability. The numerical approximation is obtained. Through the above numerical integration, the intersection probability of the first and second moving subjects within the time window is obtained.

[0044] The Gaussian distribution analytical mode refers to a method of calculating the intersection probability by integrating the joint probability density function in analytical form, based on the assumption that the trajectory error follows a Gaussian distribution. This mode eliminates the need for random sampling; instead, it directly integrates the probability density function within a spatially defined safe distance to obtain a precise intersection probability value, making it suitable for scenarios requiring high computational accuracy.

[0045] When using the Gaussian distribution analytical model, if the error of the first or second position distribution does not follow a Gaussian distribution, kernel density estimation is used to fit the distribution of the error. The specific implementation process is as follows: 1) Determine whether the error of the first or second position distribution follows a Gaussian distribution. Error refers to the deviation between the mean predicted position output by the AI ​​trajectory prediction model and the actual position of the moving subject. Collect error data between the mean predicted position and the actual position at multiple prediction time steps on the validation set to form an error sample set. Perform a normality test on the error sample set, using the Shapiro-Wilke test to calculate the test statistic and compare it with the critical value. If the test statistic is greater than the critical value, the error is determined to follow a Gaussian distribution; if the test statistic is less than or equal to the critical value, the error is determined to not follow a Gaussian distribution.

[0046] 2) When the error of the first or second position distribution is determined not to follow a Gaussian distribution, kernel density estimation is used to fit the error distribution. Kernel density estimation is a nonparametric probability density estimation method that does not require prior assumptions that the data follows a specific distribution. Instead, it fits the probability density function of the error by placing a kernel function at each error data point and summing the results. The formula for calculating the probability density function of kernel density estimation is: ,in, Indicates the location The probability density estimate at that location, Represents the error variable. This represents the total number of error data points in the error sample set. Indicates bandwidth parameter, In the error sample set, the first... The value of each error data point. This represents the kernel function. The Gaussian kernel function is chosen, and its mathematical form is: ,in, The independent variable representing the kernel function, equal , This represents the natural exponential function. Bandwidth parameter. The Silverman rule is used to determine this. The formula for calculating the Silverman rule is as follows: ,in, This represents the standard deviation estimate of the error sample set. This represents the total number of error data points in the error sample set.

[0047] 3) Replace the original Gaussian noise generation method with the error distribution fitted by kernel density estimation. After determining the probability density function of kernel density estimation, random sampling is performed from the probability density function to generate error values. The generated error values ​​are then superimposed on the mean of the predicted positions output by the AI ​​trajectory prediction model to obtain new position samples. After the above processing, when the error of the first position distribution does not follow a Gaussian distribution, the sample generation of the first position distribution uses the error distribution fitted by kernel density estimation; when the error of the second position distribution does not follow a Gaussian distribution, the sample generation of the second position distribution uses the error distribution fitted by kernel density estimation. The error distribution fitted by kernel density estimation can more accurately characterize the true probability distribution of the error, improving the accuracy of subsequent intersection probability calculations.

[0048] S4. Generate a collision avoidance control command or a warning signal based on the encounter probability, and send the collision avoidance control command to the actuator of the first moving body and / or the second moving body, or send the warning signal to the display terminal or alarm device, specifically: The probability of a collision is compared with a preset first threshold and a second threshold, where the first threshold is greater than the second threshold. When the probability of a collision is greater than or equal to the first threshold, an emergency braking command is generated as a collision avoidance control command, instructing the actuators of the first and / or second moving bodies to perform an emergency braking operation. When the probability of a collision is less than the first threshold but greater than or equal to the second threshold, a deceleration and avoidance command is generated as a collision avoidance control command, or a warning signal containing risk location and risk time information is generated and sent to the display terminal or alarm device. When the probability of a collision is less than the second threshold, no collision avoidance control command is generated and no warning signal is generated. The specific implementation process is as follows: S40, the first threshold, and the second threshold are all preset probability values, with the first threshold being greater than the second threshold. The first threshold corresponds to a high-risk level, indicating that when the probability of encounter reaches this value, a collision is highly likely between the two moving entities, requiring immediate forced braking. The second threshold corresponds to a medium-risk level, indicating that when the probability of encounter reaches this value, there is a significant risk of collision, requiring deceleration and avoidance measures or issuing a warning to the operator. The specific values ​​of the first and second thresholds are pre-calibrated based on the type of moving entity, safety level requirements, and application scenario. For satellite collision warning scenarios, the first threshold is set to 0.1%, and the second threshold is set to 0.01%; for vehicle collision avoidance scenarios, the first threshold is set to 0.5%, and the second threshold is set to 0.1%.

[0049] S41. The intersection probability is a value calculated within a time window using a Monte Carlo simulation model or a Gaussian distribution analytical model, denoted as S41. The probability of intersection With the first threshold Comparison, This represents the first threshold. When... When the risk level is determined to be high, then... If the comparison is not completed, proceed to step S43 to continue the subsequent comparison.

[0050] S42. When the probability of a collision is greater than or equal to a first threshold, an emergency braking command is generated as a collision avoidance control command. The emergency braking command is a control command containing a braking command identifier and an emergency braking acceleration value. The emergency braking acceleration value is set to the maximum braking deceleration that the actuator of the moving body can provide. For vehicles, the maximum braking deceleration is taken as the design limit value of the vehicle's braking system; for satellites, the emergency braking command corresponds to an attitude adjustment or orbital maneuver command, instructing the thrusters to apply the maximum reverse thrust. The emergency braking command also carries an execution timestamp, the value of which is the current time plus a preset command transmission and execution delay time. The generated emergency braking command is sent to the actuator of the first or second moving body. Upon receiving the command, the actuator performs an emergency braking operation according to the emergency braking acceleration value, causing the moving body to decelerate or stop in the shortest possible time, thereby avoiding a collision.

[0051] When the intersection probability is less than the first threshold, the intersection probability is compared with the second threshold. The second threshold is denoted as... , Indicates the second threshold, and The probability of intersection With the second threshold Comparison. When When the risk level is determined to be medium risk; when At that time, the current risk level is determined to be low risk.

[0052] When the probability of a collision is less than a first threshold but greater than or equal to a second threshold, a deceleration and avoidance command is generated as a collision avoidance control command, or a warning signal containing risk location and risk time information is generated and sent to the display terminal or alarm device. The deceleration and avoidance command includes a deceleration acceleration value and an avoidance direction deflection angle value. The deceleration acceleration value is less than the emergency braking acceleration value. The avoidance direction deflection angle is determined based on the relative positions of the first and second moving entities, allowing the moving entity to increase its distance from the other moving entity through moderate deceleration and minor directional adjustments. The warning signal includes risk location and risk time information. The risk location is obtained by calculating the midpoint of the two distribution mean points of the first and second position distributions at the time step with the highest probability of collision. The risk time is the moment corresponding to the time step with the highest probability of collision. The warning signal is encoded into a standard data frame format and sent to the display terminal or alarm device. Upon receiving the warning signal, the display terminal displays the risk location and risk time on the screen. Upon receiving the warning signal, the alarm device triggers an audible and visual alarm, prompting the operator to pay attention to the risk and manually take avoidance measures. In scenarios where a deceleration and avoidance command is generated, the deceleration and avoidance command is sent to the actuator of the first or second moving body, and the actuator performs deceleration and fine-tuning of direction according to the deceleration and avoidance command.

[0053] When the probability of a collision is less than the second threshold, no collision avoidance control command or warning signal is generated. In this case, it is determined that there is no foreseeable collision risk between the two moving entities within the time window, and no intervention is required. The collision probability calculation process ends, and the moving entities continue to operate in their current motion state.

[0054] Collision avoidance control commands are a general term for control commands used to guide moving entities to perform collision avoidance operations, including two forms: emergency braking commands and deceleration and avoidance commands. Collision avoidance control commands are generated by the upper-level decision-making system and sent to the actuators of the moving entity via a communication link. The actuators then perform corresponding mechanical operations according to the command content. Deceleration and avoidance commands are control commands containing deceleration acceleration values ​​and avoidance direction deflection angle values. Deceleration and avoidance commands instruct the actuators of the moving entity to perform appropriate deceleration operations and fine-tune the direction of movement to increase the relative distance with another moving entity, thereby reducing the risk of collision. The braking intensity of deceleration and avoidance commands is less than that of emergency braking commands, and they are suitable for scenarios with a medium-risk probability of encounter. Warning signals are information data frames containing risk location and risk time information. Warning signals do not directly control the actuators of the moving entity but are sent to a display terminal or alarm device to alert operators of the current collision risk. Operators then manually decide whether to take intervention measures based on the displayed information. Warning signals are suitable for scenarios with a medium-risk probability of encounter and where manual intervention is permissible. The risk location refers to the midpoint of the line connecting the average predicted positions of the first and second moving subjects at the moment when the probability of their intersection within the time window is highest. The risk location reflects the spatial location where a collision is most likely to occur and is used to assist operators in identifying risk areas. The risk time refers to the point in time corresponding to the moment when the probability of the intersection of the two moving subjects within the time window is highest. The risk time reflects the moment when a collision is most likely to occur and is used to assist operators in assessing the urgency of the risk. The actuator is a mechanical device on a moving subject used to perform physical actions such as braking, steering, or thrust control. For vehicles, the actuator includes brakes, steering gear, and throttle controllers; for satellites, the actuator includes thrusters and attitude control wheels; for drones, the actuator includes motors and control surfaces. After receiving collision avoidance control commands, the actuator converts the acceleration or angle values ​​in the commands into actual mechanical actions, thereby changing the motion state of the moving subject.

[0055] Optionally, the above technical solution also includes: S5. For scenarios with multiple moving subjects, the multiple moving subjects are decomposed into multiple pairs of moving subjects. For each pair of moving subjects, the steps of initial motion parameter acquisition, trajectory distribution prediction, and intersection probability calculation are performed separately, and the intersection probability of each pair of moving subjects is output. Alternatively, the joint intersection probability of all moving subjects is calculated by combining them through a preset hazard domain simplification strategy. Based on the intersection probability or joint intersection probability of each pair of moving subjects, collision avoidance control commands or warning signals are generated. The collision avoidance control commands are sent to the actuators of the corresponding moving subjects, or the warning signals are sent to the display terminal or alarm device.

[0056] For scenarios with multiple moving subjects, the multiple moving subjects are decomposed into multiple pairs of moving subjects. For each pair of moving subjects, the steps of initial motion parameter acquisition, trajectory distribution prediction, and intersection probability calculation are performed, and the intersection probability of each pair of moving subjects is output. The specific implementation process is as follows: S50. When the total number of moving entities participating in the collision risk assessment is greater than or equal to three, the current scenario is determined to be a multi-moving-entity scenario. In a multi-moving-entity scenario, there may be collision risk between any two moving entities. It is necessary to calculate the collision probability for each pair of all moving entities in order to comprehensively assess the collision risk distribution in the entire scenario.

[0057] S51. A pair of moving subjects refers to an ordered pairing formed by arbitrarily selecting two moving subjects from all moving subjects. The total number of pairs of moving subjects is calculated using the combination number formula. If the total number of moving subjects is... , If the total number of moving subjects is represented by , then the total number of pairs of moving subjects is equal to the number of moving subjects from . The number of ways to combine two objects from a set of moving objects is denoted as . ,Right now Multiply Divide by two. Iterate through all moving subjects in sequence, and combine every two different moving subjects into a moving subject pair. Each moving subject pair contains the identity identifiers of the first moving subject and the second moving subject. The identity identifiers are used to distinguish different moving subjects and associate them with their respective acquired initial motion parameters.

[0058] S52. Based on the identity identifier recorded in the centering record of the moving subjects, obtain the initial position, initial velocity, and initial direction of motion of the first moving subject and the second moving subject. The method for obtaining the initial motion parameters is the same as that for a single scene. The position, velocity, and direction of motion data of each moving subject are collected through a global navigation satellite system receiver, inertial measurement unit, radar, or optical sensor, and the initial motion parameters in a unified coordinate system are obtained after preprocessing.

[0059] S53. Input the initial motion parameters of the first moving subject in each moving subject pair into a pre-trained AI trajectory prediction model to obtain the first position distribution of the first moving subject within the time window; input the initial motion parameters of the second moving subject into the same AI trajectory prediction model to obtain the second position distribution of the second moving subject within the time window. For scenarios involving multiple moving subjects, different moving subject pairs share the same AI trajectory prediction model, and the model parameters do not change due to the switching of moving subject pairs. Each moving subject pair independently completes its trajectory distribution prediction, and the prediction processes between different moving subject pairs do not interfere with each other.

[0060] S54. For each pair of moving subjects, the intersection probability within the time window is calculated using either a Monte Carlo simulation mode or a Gaussian distribution analytical mode, based on a preset mode selection. In the Monte Carlo simulation mode, position samples are collected from the first and second position distributions respectively. The distance between samples is calculated at each time step, and the number of intersections where the distance between samples is less than or equal to the safe distance threshold is counted. The intersection probability is obtained by dividing the number of intersections by the product of the total number of samples and the total number of time steps. In the Gaussian distribution analytical mode, the first Gaussian parameter of the first position distribution and the second Gaussian parameter of the second position distribution are extracted. The Gaussian distribution of the relative positions is calculated, and the intersection probability is obtained by numerically integrating the Gaussian distribution of the relative positions within the area defined by the safe distance threshold. The intersection probability calculation process for each pair of moving subjects is independent. The same calculation mode can be selected, or different calculation modes can be selected according to the specific scenario requirements of each pair of moving subjects.

[0061] S55. Associate and store the identity identifier of each moving entity pair with the calculated intersection probability value for that moving entity pair to form a intersection probability list. Each item in the intersection probability list contains the identity identifier of the first moving entity in the moving entity pair, the identity identifier of the second moving entity, and the corresponding intersection probability value. Output the intersection probability list for use in generating subsequent collision avoidance control commands or warning signals.

[0062] The joint intersection probability of all moving subjects is calculated by merging them using a pre-defined danger domain simplification strategy. This includes: determining the mean point of the position distribution of all moving subjects at each time step within the time window; calculating the convex hull boundary of all mean points; extending the convex hull boundary outwards by a safety distance threshold to obtain the danger domain; then calculating the ratio of the number of position samples of each moving subject falling within the danger domain to the total number of samples; and taking the maximum value of the ratio for all moving subjects as the joint intersection probability. The specific implementation process is as follows: 1) Determine the mean point of the position distribution of all moving subjects at each time step within the time window, where the time window includes the total number of time steps. Discrete time steps This represents the total number of time steps. For each time step, all moving subjects are traversed, and the mean vector of the position distribution for each moving subject is extracted. The x and y coordinates of the mean vector of the position distribution are denoted as... and ,in An index representing the moving subject. The value range is from the first moving subject to the second. One moving entity, Indicates the total number of moving entities; Indicates the index of the time step. The value range is from the first time step to the second time step. One time step; Indicates the first The moving subject in the first The x-coordinate of the mean vector of the location distribution at each time step Indicates the first The moving subject in the first The ordinate of the mean vector of position distribution at each time step. The coordinates corresponding to the mean vector of all moving subjects at each time step are taken as the set of mean points for that time step. The set of mean points at each time step contains There are 1 coordinate point, and the coordinate point is in the form of 1. , Take all the first to the second Each moving entity.

[0063] 2) Calculate the convex hull boundary of all mean points at each time step. The convex hull boundary is the smallest convex polygon boundary that can contain all mean points within that time step. For the ... At each time step, the obtained Coordinates of the mean points As the input point set, the monotonic chain convex hull algorithm is used to calculate the convex hull boundary. The execution process of the monotonic chain convex hull algorithm is as follows: Sort the input point set according to the x-coordinate in ascending order, and if the x-coordinates are the same, sort them according to the y-coordinate in ascending order; construct the upper convex hull chain and the lower convex hull chain respectively. The upper convex hull chain starts from the point with the smallest x-coordinate and connects sequentially to form the upper boundary in a counterclockwise direction. The lower convex hull chain starts from the point with the largest x-coordinate and connects sequentially to form the lower boundary in a counterclockwise direction; merge the upper convex hull chain and the lower convex hull chain, remove the first and last duplicate points to obtain a closed convex polygon. The boundary of this convex polygon is the convex hull boundary. The convex hull boundary is represented in the form of a vertex sequence, where the line segments between adjacent vertices in the vertex sequence constitute the convex hull boundary. The above convex hull boundary calculation is performed for each time step to obtain the convex hull boundary for each time step.

[0064] 3) Extend the convex hull boundary outwards by a safety distance threshold to obtain the danger zone. The safety distance threshold is denoted as... , This represents the safety distance threshold. For each time step's convex hull boundary, each edge on the convex hull boundary is shifted outward by a distance. The translation direction is perpendicular to the edge and points out of the convex hull; extend all translated edges and find their intersections to form a new closed polygon boundary; for adjacent vertices on the new closed polygon boundary, a safe distance threshold is used. Using the radius as the radius and the adjacent vertex as the center, draw an arc connecting the two edges, making the corners of the new boundary smooth arc transitions. The resulting expanded closed region is the danger zone. The minimum distance between the outer boundary of the danger zone and the original convex hull boundary in all directions is equal to the safety distance threshold. Perform an outward expansion operation on the convex hull boundary for each time step to obtain the danger zone for each time step.

[0065] 4) Calculate the ratio of the number of position samples of each moving subject falling into the danger zone to the total number of samples. The total number of samples is denoted as... , This represents the total number of samples. For each time step and each moving subject, the total number of samples is collected from the position distribution of that moving subject at that time step. There are 3 location samples, each a two-dimensional coordinate point. Each location sample is individually determined to be within the danger zone at that time step. The determination method uses a ray-mapping approach: a ray is drawn from the location sample point in any direction, and the number of intersections between the ray and the danger zone boundary is counted. If the number of intersections is odd, the location sample point is considered to be inside the danger zone; if the number of intersections is even, the location sample point is considered to be outside the danger zone. The number of samples falling inside the danger zone is counted and denoted as . , Indicates the first The moving subject in the first The number of samples falling within the danger zone at each time step. Divide by the total number of samples , obtained the The moving subject in the first The proportion of time steps falling into the danger zone is denoted as . , Perform the above calculations for each time step and each moving subject to obtain the ratio values ​​for all time steps and all moving subjects. .

[0066] 5) Take the maximum value of the ratios of all moving subjects as the joint intersection probability, and find the ratio value within the range of all moving subjects and all time steps. The maximum value in, that is, for all From the first to the... Individual, All From the first to the... Individual Take the maximum value. Use the maximum value as the joint intersection probability, denoted as... , The joint rendezvous probability reflects the highest rendezvous risk level faced by any time step and any moving subject in the overall scenario consisting of all moving subjects within the time window. This value is used to determine the risk level of the overall scenario when generating collision avoidance control commands or warning signals.

[0067] The mean point of the position distribution refers to the expected center of all possible positions of the moving subject at a given time step, determined by the predicted position mean vector output by the AI ​​trajectory prediction model. The mean point is a two-dimensional coordinate system containing horizontal and vertical coordinates, representing the most likely spatial position of the moving subject at a specified time. Each moving subject has a unique mean point of position distribution at each time step. The convex hull boundary is the smallest convex polygon boundary containing the mean point of the position distribution of all moving subjects within a given time step. The convex hull boundary is convex, meaning that the line connecting any two points on the boundary lies entirely inside or on the boundary. The convex hull boundary encloses all mean points within or on its boundary, defining the overall spatial range occupied by all moving subjects at a given time step. The danger zone is the extended region formed by expanding the convex hull boundary outwards by a safety distance threshold. The danger zone uniformly expands the original convex hull boundary outwards by the safety distance threshold in all directions, with the distance between the expanded boundary and the original boundary everywhere equal to the safety distance threshold. The danger zone characterizes the overall spatial range occupied and influenced by all moving subjects at a given time step, considering the safety interval. Joint rendezvous probability refers to the overall risk index obtained by maximizing the proportion of each moving entity's hazard domain falling into the hazard domain across all moving entities and all time steps in a multi-moving entity scenario. The joint rendezvous probability reflects the maximum rendezvous risk level faced by any moving entity in the entire scenario as a single numerical value, simplifying risk assessment and early warning decisions in multi-moving entity scenarios.

[0068] The collision avoidance control command or warning signal is generated based on the intersection probability or joint intersection probability of each moving object pair. The collision avoidance control command is sent to the actuator of the corresponding moving object, or the warning signal is sent to the display terminal or alarm device. The specific implementation process is as follows: 1) When a multi-moving-subject scenario uses a moving-subject pair decomposition method, the input is a list of intersection probabilities for each moving-subject pair. Each item in the list contains the identifiers of the first and second moving-subjects in the pair, along with their corresponding intersection probability values. When a multi-moving-subject scenario uses a hazard domain simplification strategy, the input is a joint intersection probability, which reflects the maximum intersection risk level faced by any moving-subject in the entire scenario as a single value. Based on the calculation strategy used in the actual scenario, select the corresponding input and proceed with the subsequent steps.

[0069] 2) When the input is the intersection probability of each moving subject pair, a threshold comparison and command generation operation are performed on the intersection probability of each moving subject pair. The first threshold and the second threshold are preset probability values, with the first threshold being greater than the second threshold. The intersection probability of the current moving subject pair is compared with the first threshold. When the intersection probability is greater than or equal to the first threshold, an emergency braking command is generated as a collision avoidance control command. The emergency braking command contains the identification identifiers of the first and second moving subjects in the current moving subject pair, as well as the emergency braking acceleration value and the current timestamp. This command instructs the actuator of the moving subject corresponding to the identification identifier to perform an emergency braking operation. When the probability of a collision is less than a first threshold, the probability is compared with a second threshold. When the probability is less than the first threshold but greater than or equal to the second threshold, a deceleration and avoidance command is generated as a collision avoidance control command, or a warning signal containing risk location and risk time information is generated. The deceleration and avoidance command includes the identification of the two moving objects in the current moving object pair, the deceleration acceleration value, and the avoidance direction deflection angle value. The warning signal includes the identification of the two moving objects in the current moving object pair and the calculated risk location and risk time information. When the probability of a collision is less than the second threshold, no collision avoidance control command or warning signal is generated for that moving object pair. The above threshold comparison and command generation operations are performed one by one for each moving object pair in the collision probability list, generating a corresponding collision avoidance control command or warning signal for each moving object pair, or determining that no command or signal is generated.

[0070] 3) When the input is based on the joint intersection probability, a threshold comparison and command generation operation is performed on the joint intersection probability. The joint intersection probability is compared with a first threshold. When the joint intersection probability is greater than or equal to the first threshold, the overall scene consisting of all moving objects is determined to be at a high-risk level, and a global emergency braking command is generated as a collision avoidance control command. The global emergency braking command contains a list of the identities of all moving objects and instructs the actuators of all moving objects to simultaneously perform emergency braking operations. When the probability of joint encounter is less than the first threshold, the probability of joint encounter is compared with the second threshold. When the probability of joint encounter is less than the first threshold but greater than or equal to the second threshold, the overall scenario is determined to be at a medium risk level. A global deceleration and avoidance command is generated as a collision avoidance control command, or a warning signal containing the global risk location and global risk time is generated. The global deceleration and avoidance command contains a list of the identities of all moving entities and instructs the actuators of all moving entities to simultaneously perform deceleration and avoidance operations. The global risk location is obtained by taking the geometric center of the average predicted location of all moving entities at the moment of maximum joint encounter probability. The global risk time is the time corresponding to the time step with the maximum joint encounter probability. When the probability of joint encounter is less than the second threshold, the overall scenario is determined to be at a low risk level. No collision avoidance control command or warning signal is generated.

[0071] 4) Transmitting Collision Avoidance Control Commands. Each generated collision avoidance control command is encapsulated into a standard data frame format. The data frame contains a command type identifier, the target moving entity's identification, control parameter values, and an execution timestamp. Based on the target moving entity's identification carried in the collision avoidance control command, the command data frame is transmitted to the corresponding target moving entity via a wireless communication link. Upon receiving the command data frame, the target moving entity's actuator decodes the data frame, extracts the control parameters, and performs the corresponding emergency braking or deceleration avoidance operation according to the command content.

[0072] 5) Each generated warning signal is encapsulated into a standard data frame format. The data frame contains a signal type identifier, the identity of the target moving object, and information on the risk location and time. Based on the target moving object's identity carried in the warning signal, the signal data frame is sent to the corresponding display terminal or alarm device via a wired or wireless communication link. Upon receiving the warning signal, the display terminal displays the risk location and time information in graphical and textual form on the screen interface; upon receiving the warning signal, the alarm device triggers an audible alarm and flashing lights to alert the operator of the current collision risk.

[0073] The corresponding moving subject refers to one or more moving subjects specified in the collision avoidance control command or warning signal that need to perform the corresponding operation. In the moving subject pair decomposition method, the corresponding moving subject is the two moving subjects in the moving subject pair targeted by the intersection probability calculation; in the hazard domain simplification strategy, the corresponding moving subject is all moving subjects. The identification of the corresponding moving subject is explicitly recorded in the command data frame to ensure that the command is sent only to the correct receiver. The global emergency braking command is a collision avoidance control command containing a list of identifications of all moving subjects, instructing all moving subjects to simultaneously perform braking operations at the maximum available deceleration of their respective actuators. The global emergency braking command is generated when the overall risk level of the multi-moving subject scenario reaches high risk, and is used to implement unified emergency intervention for the entire scenario. The global deceleration and avoidance command is a collision avoidance control command containing a list of identifications of all moving subjects, instructing all moving subjects to simultaneously perform moderate deceleration operations and directional fine-tuning. The global deceleration and avoidance command is generated when the overall risk level of the multi-moving subject scenario is medium risk, and is used to implement unified moderate intervention for the entire scenario. The global risk location refers to the geometric center coordinates formed by the average predicted positions of all moving entities at the time step with the highest probability of joint intersection. The global risk location reflects the overall spatial center region at the moment of highest collision risk in a multi-moving-entity scenario, facilitating operators' assessment of the spatial distribution of overall risk. The global risk time refers to the moment corresponding to the time step with the highest probability of joint intersection, reflecting the time point of highest overall risk in a multi-moving-entity scenario, facilitating operators' assessment of the urgency of the risk.

[0074] This invention first obtains the initial position, initial velocity, and initial direction of motion of each moving subject. After coordinate system transformation and timestamp alignment, these are input into a pre-trained Long Short-Term Memory (LSTM) network model. This model employs a two-layer, bidirectional LSM structure, using the position, velocity, and direction sequences of historical time steps as input to predict the mean trajectory position at each time step within a future time window. Gaussian noise is added to the mean to quantify prediction uncertainty, thus obtaining the position distribution of each moving subject at each time step. When the error does not follow a Gaussian distribution, kernel density estimation is used to fit the error distribution, and position samples are generated from it.

[0075] The intersection probability calculation offers two switchable modes. In Monte Carlo simulation mode, position samples equal to the total number of samples are collected from the position distributions of the two subjects. Samples are paired by index, and the Euclidean distance between each pair is calculated. The number of intersections within a safe distance threshold is counted, and the intersection probability is obtained by dividing the total number of intersections by the product of the total number of samples and the total number of time steps. In Gaussian distribution analytical mode, the mean vectors and covariance matrices of the two position distributions are extracted. The mean difference and covariance sum of the relative positions are calculated. The intersection probability is obtained by numerically integrating the Gaussian probability density of the relative positions within a circular region defined by the safe distance threshold. In Monte Carlo simulation mode, the total number of samples is dynamically adjusted according to preset accuracy requirements. Accuracy is quantified by relative error limits. The standard deviation of the estimated value is inversely proportional to the square root of the number of samples. The number of samples is adjusted between 5,000 and 20,000, and the step size is gradually increased when the accuracy requirement is not met. In analytical mode, if the error does not follow a Gaussian distribution, kernel density estimation is used. The error probability density is fitted with the bandwidth determined by the Gaussian kernel function and Silverman rule, and the original Gaussian noise is sampled from it to replace it in order to adapt to the non-Gaussian error shape.

[0076] For scenarios with multiple moving subjects, the system can be decomposed into multiple pairs of moving subjects. Each pair independently performs initial parameter acquisition, trajectory prediction, and intersection probability calculation, outputting a list of intersection probabilities for each pair. Alternatively, a hazard domain simplification strategy can be adopted. First, the mean point of the distribution of all subject positions at each time step is determined, and its convex hull boundary is calculated. The boundary is then extended outward by a safety distance threshold to form a hazard domain. The percentage of each subject position sample falling into the hazard domain is then calculated, and the maximum percentage across all subjects and all time steps is taken as the joint intersection probability. Based on the intersection probability or joint intersection probability of each pair of moving subjects, it is compared with a first threshold and a second threshold. If the probability is higher than or equal to the first threshold, an emergency braking command is generated, including the target identification and maximum braking deceleration. If the probability is between the two thresholds, a deceleration and avoidance command or a warning signal including the risk location and risk time is generated. If the probability is lower than the second threshold, no command or signal is generated. All commands are encapsulated into standard data frames and transmitted to the corresponding actuators via a wireless link. Warning signals are sent to the display terminal or alarm device. The algorithm functions are encapsulated into a standardized interface, defining input parameters and output data formats, which can be embedded into satellite collision warning or intelligent transportation systems to achieve real-time calculation of rendezvous probability and automated risk warning.

[0077] The beneficial effects of this invention include: ① By outputting the trajectory distribution through an AI trajectory prediction model and adding Gaussian noise to the predicted position to quantify uncertainty, the intersection probability calculation fully considers the statistical characteristics of the prediction error, resulting in calculation results that are more consistent with actual scenarios and effectively reducing the risk of misjudgment and missed judgment. ② It provides two switchable calculation schemes: Monte Carlo simulation mode and Gaussian distribution analytical mode. The former is suitable for general scenarios, while the latter is suitable for high-precision scenarios. The choice can be flexible according to different accuracy and real-time requirements, and it is applicable to various moving subjects such as satellites, vehicles, and drones. ③ Through a dynamic adjustment strategy for the number of samples, the total number of samples is automatically adjusted between 5,000 and 20,000, balancing calculation accuracy and speed. For scenarios with multiple moving subjects, a strategy of decomposing moving subjects or simplifying the danger zone is adopted to reduce computational complexity and meet the real-time requirements of high-frequency scenarios. ④ From initial motion parameter acquisition, trajectory prediction, intersection probability calculation to collision avoidance control commands or warning signals generation, the entire process is automated without manual intervention. The algorithm functions are encapsulated into a standardized interface, which can be directly embedded into existing systems such as satellite collision warning and intelligent traffic management, resulting in low implementation costs, high practicality, and effectively improving the intelligence level of the system.

[0078] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given by the present invention. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation. The scheme after adjusting the order is also within the protection scope of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0079] like Figure 2 As shown in the figure, an AI-based moving subject intersection probability determination and collision avoidance warning system according to an embodiment of the present invention includes a motion parameter acquisition module, a prediction addition module, an intersection probability calculation module and a collision avoidance warning module; The motion parameter acquisition module is used to: acquire the initial motion parameters of the first moving body and the second moving body; The prediction addition module is used to: input the initial motion parameters into the pre-trained AI trajectory prediction model, predict the trajectory distribution of the first moving subject and the second moving subject within the time window respectively, and add Gaussian noise to the trajectory distribution to quantify the prediction uncertainty, so as to obtain the first position distribution of the first moving subject and the second position distribution of the second moving subject. The intersection probability calculation module is used to: obtain the intersection probability of the first moving subject and the second moving subject based on the first position distribution and the second position distribution; The collision avoidance warning module is used to: generate collision avoidance control commands or generate warning signals based on the probability of encounter, and send the collision avoidance control commands to the actuators of the first moving body and / or the second moving body, or send the warning signals to the display terminal or alarm device.

[0080] Optionally, in the above technical solution, the intersection probability calculation module is specifically used for: when using the Monte Carlo simulation mode, collecting multiple sets of position samples from the first position distribution and the second position distribution respectively, traversing each time step, calculating the sample distance between the first moving subject and the second moving subject at each time step, counting the number of intersections where the sample distance is less than or equal to the safe distance threshold, and dividing the number of intersections by the product of the total number of samples and the total number of time steps to obtain the intersection probability; or, when using the Gaussian distribution analytical mode, extracting the first Gaussian parameter of the first position distribution and the second Gaussian parameter of the second position distribution, calculating the Gaussian distribution of the relative position based on the first Gaussian parameter and the second Gaussian parameter, and obtaining the intersection probability by numerically integrating the Gaussian distribution of the relative position within the area limited by the safe distance threshold.

[0081] Optionally, in the above technical solution, the intersection probability calculation module is also used to: dynamically adjust the total number of samples according to the preset accuracy requirements when using the Monte Carlo simulation mode; and / or, when using the Gaussian distribution analytical mode, when the error of the first position distribution or the second position distribution does not follow the Gaussian distribution, use kernel density estimation to fit the distribution of the error.

[0082] Optionally, the above technical solution also includes a decomposition and invocation module, which is used to: for a multi-moving-subject scenario, decompose the multi-moving-subject into multiple moving-subject pairs, and invoke the motion parameter acquisition module, the prediction addition module, and the intersection probability calculation module to perform the steps of initial motion parameter acquisition, trajectory distribution prediction, and intersection probability calculation for each moving-subject pair, and output the intersection probability of each moving-subject pair; or, invoke the intersection probability calculation module to merge and calculate the joint intersection probability of all moving-subjects through a preset hazard domain simplification strategy; and invoke the collision avoidance warning module to generate collision avoidance control commands or warning signals based on the intersection probability or joint intersection probability of each moving-subject pair, and send the collision avoidance control commands to the actuators of the corresponding moving-subjects, or send the warning signals to the display terminal or alarm device.

[0083] Optionally, in the above technical solution, the intersection probability calculation module is specifically used to: determine the mean point of the position distribution of all moving subjects at each time step within the time window, calculate the convex hull boundary of all mean points, extend the convex hull boundary outwards by a safety distance threshold to obtain the danger zone, then calculate the ratio of the number of position samples of each moving subject falling into the danger zone to the total number of samples, and take the maximum value of the ratio of all moving subjects as the joint intersection probability.

[0084] Optionally, in the above technical solution, the collision avoidance warning module is specifically used to: compare the meeting probability with a preset first threshold and a second threshold, wherein the first threshold is greater than the second threshold; when the meeting probability is greater than or equal to the first threshold, generate an emergency braking command as a collision avoidance control command, wherein the emergency braking command instructs the actuators of the first moving body and / or the second moving body to perform an emergency braking operation; when the meeting probability is less than the first threshold but greater than or equal to the second threshold, generate a deceleration and avoidance command as a collision avoidance control command, or generate a warning signal containing risk location and risk time information and send it to a display terminal or alarm device.

[0085] It should be noted that the beneficial effects of the AI-based moving subject collision probability determination and collision avoidance warning system provided in the above embodiments are the same as those of the AI-based moving subject collision probability determination and collision avoidance warning method described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0086] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned AI-based methods for determining the probability of intersection of moving subjects and for collision avoidance warning.

[0087] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned AI-based methods for determining the probability of intersection of moving subjects and for collision avoidance warning.

[0088] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0089] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for determining the probability of collision and providing collision avoidance warning based on AI for moving subjects, characterized in that, include: Obtain the initial motion parameters of the first and second moving subjects; The initial motion parameters are input into a pre-trained AI trajectory prediction model to predict the trajectory distributions of the first moving subject and the second moving subject within the time window, respectively. Gaussian noise is added to the trajectory distributions to quantify the prediction uncertainty, thereby obtaining the first position distribution of the first moving subject and the second position distribution of the second moving subject. Based on the first position distribution and the second position distribution, the intersection probability of the first moving subject and the second moving subject is obtained; Based on the intersection probability, a collision avoidance control command or a warning signal is generated, and the collision avoidance control command is sent to the actuator of the first moving body and / or the second moving body, or the warning signal is sent to the display terminal or alarm device.

2. The method for determining the probability of collision and providing collision avoidance warning based on AI for moving subjects according to claim 1, characterized in that, Based on the first position distribution and the second position distribution, the intersection probability of the first moving subject and the second moving subject is obtained, including: When using the Monte Carlo simulation mode, multiple sets of position samples are collected from the first position distribution and the second position distribution, respectively. Each time step is iterated, and the distance between the samples of the first moving subject and the second moving subject at each time step is calculated. The number of intersections where the distance between the samples is less than or equal to the safe distance threshold is counted. The intersection probability is obtained by dividing the number of intersections by the product of the total number of samples and the total number of time steps. Alternatively, when using the Gaussian distribution analytical mode, the first Gaussian parameter of the first position distribution and the second Gaussian parameter of the second position distribution are extracted. The Gaussian distribution of the relative position is calculated based on the first Gaussian parameter and the second Gaussian parameter. The intersection probability is obtained by numerically integrating the Gaussian distribution of the relative position within the area defined by the safe distance threshold.

3. The method for determining the probability of collision and providing collision avoidance warning based on AI for moving subjects according to claim 2, characterized in that, When the Monte Carlo simulation mode is used, the total number of samples is dynamically adjusted according to the preset accuracy requirements; and / or, when the Gaussian distribution analytical mode is used, if the error of the first position distribution or the second position distribution does not follow a Gaussian distribution, kernel density estimation is used to fit the distribution of the error.

4. A method for determining the probability of collision and providing early warning of collision avoidance for moving subjects based on AI, as described in any one of claims 1 to 3, characterized in that, Also includes: For scenarios with multiple moving entities, the multiple moving entities are decomposed into multiple pairs of moving entities. For each pair of moving entities, the steps of initial motion parameter acquisition, trajectory distribution prediction, and intersection probability calculation are performed, and the intersection probability of each pair of moving entities is output. Alternatively, the joint intersection probability of all moving entities is calculated by combining them through a preset hazard domain simplification strategy. Collision avoidance control commands or warning signals are generated based on the intersection probability of each pair of moving entities or the joint intersection probability. The collision avoidance control commands are sent to the actuators of the corresponding moving entities, or the warning signals are sent to the display terminal or alarm device.

5. An AI-based system for determining the probability of collisions between moving subjects and for collision avoidance warning, characterized in that, It includes a motion parameter acquisition module, a prediction addition module, a intersection probability calculation module, and a collision avoidance warning module; The motion parameter acquisition module is used to: acquire the initial motion parameters of the first moving body and the second moving body; The prediction addition module is used to: input the initial motion parameters into a pre-trained AI trajectory prediction model, predict the trajectory distributions of the first moving subject and the second moving subject within the time window, and add Gaussian noise to the trajectory distributions to quantify the prediction uncertainty, thereby obtaining the first position distribution of the first moving subject and the second position distribution of the second moving subject; The intersection probability calculation module is used to: obtain the intersection probability of the first moving subject and the second moving subject based on the first position distribution and the second position distribution; The collision avoidance warning module is used to: generate a collision avoidance control command or a warning signal based on the meeting probability, and send the collision avoidance control command to the actuator of the first moving body and / or the second moving body, or send the warning signal to the display terminal or alarm device.

6. The AI-based system for determining the probability of collision and providing early warning of collision avoidance for moving subjects according to claim 5, characterized in that, The intersection probability calculation module is specifically used for: when using the Monte Carlo simulation mode, collecting multiple sets of position samples from the first position distribution and the second position distribution respectively, traversing each time step, calculating the sample distance between the first moving subject and the second moving subject at each time step, counting the number of intersections where the sample distance is less than or equal to the safe distance threshold, and dividing the number of intersections by the product of the total number of samples and the total number of time steps to obtain the intersection probability; or, when using the Gaussian distribution analytical mode, extracting the first Gaussian parameter of the first position distribution and the second Gaussian parameter of the second position distribution, calculating the Gaussian distribution of the relative position based on the first Gaussian parameter and the second Gaussian parameter, and obtaining the intersection probability by numerically integrating the Gaussian distribution of the relative position within the area limited by the safe distance threshold.

7. The AI-based system for determining the probability of encounters between moving subjects and for collision avoidance early warning, as described in claim 6, is characterized in that... The intersection probability calculation module is also used to: dynamically adjust the total number of samples according to the preset accuracy requirements when the Monte Carlo simulation mode is used; and / or, when the Gaussian distribution analytical mode is used, when the error of the first position distribution or the second position distribution does not follow a Gaussian distribution, use kernel density estimation to fit the distribution of the error.

8. A motion subject collision probability determination and collision avoidance early warning system based on AI according to any one of claims 5 to 7, characterized in that, It also includes a decomposition and invocation module, which is used to: for multi-moving-subject scenarios, decompose the multi-moving-subject into multiple moving-subject pairs, and invoke the motion parameter acquisition module, the prediction and addition module, and the intersection probability calculation module to perform the steps of initial motion parameter acquisition, trajectory distribution prediction, and intersection probability calculation for each moving-subject pair, and output the intersection probability of each moving-subject pair; or, invoke the intersection probability calculation module to merge and calculate the joint intersection probability of all moving-subjects through a preset hazard domain simplification strategy; and invoke the collision avoidance warning module to generate collision avoidance control commands or warning signals based on the intersection probability of each moving-subject pair or the joint intersection probability, and send the collision avoidance control commands to the actuators of the corresponding moving-subjects, or send the warning signals to the display terminal or alarm device.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the AI-based method for determining the probability of intersection of moving subjects and for collision avoidance warning as described in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the AI-based method for determining the probability of intersection of moving subjects and for collision avoidance warning, as described in any one of claims 1 to 4.

Images